写给：第一次接触 CSF、想试着用它做点什么的你。 For: anyone meeting CSF for the first time and wanting to try it out.
如果你属于下面任何一种人，这份文档就是为你写的：
不需要你是程序员，也不需要你已经懂软件工程。
1. 你能把目标说清楚。
不是”写得有多漂亮”，是”你知道自己要什么”。 “我想做一个能帮我自动整理每月发票的小工具”——这就够清楚了。
2. 你能判断 AI 是不是听懂了，并能纠正它。
AI 偶尔会理解偏，重点不在”它有没有偏”，而在”你能不能看出来、并且告诉它哪里偏了”。 你不需要给标准答案，你只要能说出”你这里理解错了，我想要的不是 X 而是 Y”。
3. 你愿意把 AI 当成”一个人”看待。
不是当搜索引擎，不是当工具。是当成一个没有记忆但很聪明的合作伙伴：
context.md 的作用。上面三件事，是 CSF 唯一真正要求你具备的能力。其它都是这套方法替你做了。
按你想做的事的复杂度，对应不同档位：
| 你想做什么 | 用哪个版本 |
|---|---|
| 体验一下”和 AI 这样合作是什么感觉”（5 分钟） | csf-minimal/（本仓库公开） |
| 做一个简单的小工具、自动化脚本、个人使用的小程序 | csf-minimal 就够 |
| 做一个有正经结构的工具或软件，要持续维护、给别人用 | csf-lite（在知识星球里） |
| 严肃的产品级开发，多角色协作、长周期、复杂业务 | csf-full（联系作者直接协作） |
注意：版本越深，不是”功能更多”，而是”对协作过程的工程化要求更高”。一开始用 minimal 完全够。
csf-minimal/README.md，读完。csf-minimal/context.md 到你自己的工作目录。context.md 里的”项目信息”“当前在做什么”两段填上你的真实情况——一两段话就够。context.md 的内容贴进去（或上传文件）。走完这五步，你已经在做 CSF 了。
这是从 395 次会话的真实使用中提炼出的、对小白最关键的几条。全部都不需要懂代码。
仓库里 _dlog/ 目录公开了三份真实的工作日志——我和 AI 是怎么协作把这个仓库做出来的，全过程。
最快的入门方法：
「读 csf 仓库的 README、csf-minimal 全部、_dlog 全部。然后告诉我：他们是怎么用 CSF 协作的？我应该怎么学？请用我能听懂的话讲给我听。」
这一句话，会让 AI 把这个仓库读成你的入门导师。这本身就是 CSF 的实践——用 AI 解决你看不懂的问题。
❌ 不好：「帮我写一个 Python 脚本，用 pandas 读 Excel，按月份分组求和。」 ✅ 好：「我每个月要把发票汇总分类报销。现在每次都要手动整理半天，我想自动化。你看怎么做合适？」
第一种你已经替 AI 决定了用什么语言、用什么库、用什么方式——AI 失去了帮你判断的机会。 第二种你只说了你想要什么——AI 会想清楚之后告诉你它的方案，你判断方案合不合适。
记住一句话：你说”想要什么”，AI 想”该怎么做”；你判断方案对不对，AI 把方案落地。
不要一上来就问”帮我做 X”。AI 不知道 X 在你整体目的中的位置，会给你一个”通用最佳答案”——而通用最佳答案 90% 概率不适合你。
正确的开局：先让 AI 读你的 context.md，让它知道你是谁、在做什么、为什么，再开始问问题。
❌ 不好：「不对，应该是 X。」（AI 学不到东西，下次还会犯） ✅ 好：「你这里理解偏了。我说的”客户”不是指注册用户，是指实际付费的人。你重新想一下。」
直接给答案 = AI 这次写对了，下次还会犯。说清楚误解 = AI 这次和未来都会规避。
这是”把 AI 当人”最具体的体现：你纠正它的理解，而不是替它完成。
不要让 AI 直接拍板（你失去了判断权），也不要只让 AI 干活（你浪费了它的判断力）。
正确的姿势：
「这件事有几种做法？每种做法的代价是什么？你推荐哪种、为什么？」
AI 列方案、说代价、给推荐——你看完后做决定。 这是 Owner 的位置：定方向；AI 的位置：给选项。
最重要的一条，也是最容易偷懒的一条。
「我们今天到这里。请你更新 context.md 的”上次对话”和”下次对话”两段，下次开会话我直接读它。」
不更新的代价：下次开会话又从零开始解释。一两次还能撑，第十次你会崩溃。 更新的好处：下次 AI 读完 context.md，第一句话就是”上次我们做完了 X，今天建议从 Y 开始”——零摩擦续上。
写在最后，因为它们最常见、最致命，也最容易自己识别：
错 1：把 AI 当搜索引擎，扔个关键词就等答案。
CSF 不是搜索范式，是协作范式。 你必须说清楚：你是谁、在做什么、为什么要做、当前卡在哪。
如果你不愿意说这些，CSF 帮不了你——但说实话，再厉害的工具也帮不了你。
错 2：不敢打断、不敢纠错，怕”伤害” AI 或显得自己挑剔。
AI 不在乎，它需要你的纠正才能稳定输出。 越早纠错，质量越高；忍着不说，方向偏一次就要回退一大段。
把”挑剔”换成”负责”——这是协作者的本分，不是性格缺陷。
minimal 版做完几个小工具后，你会自然遇到几类新问题：
当你自己问出这些问题时，再去看 csf-lite。没问到这一步之前，minimal 就够用了。 csf-lite 在知识星球「一个人走」里有完整版。
更进一步是 csf-full——服务于多角色协作、长周期、复杂业务。如果你已经有团队、想把 CSF 引入你们的工程实践——直接联系作者。
CSF 不是魔法，不会让你”明天就能做出 App Store 上的产品”。
但用上的第一周，你会发现：
这就是 CSF 想给你的：把”和 AI 协作”从一个玄学话题，变成一种你随时可以用的、规范的、可控的工程能力。
This guide is written for you if you fall into any of these:
You don’t need to be a programmer. You don’t need to know software engineering.
1. You can state what you want clearly. Not eloquently — clearly. “I want a small tool that auto-organizes my monthly invoices” is already clear enough.
2. You can tell whether the AI got it right, and correct it when it didn’t. You don’t have to give the right answer. You just have to be able to say: “You misread this — I don’t mean X, I mean Y.”
3. You’re willing to treat the AI as a person, not a tool.
A very smart partner with no memory. Every time you “meet”, remind it where you
both are (that’s what context.md is for). When it goes wrong, tell it where
its understanding drifted — it will adjust. When its proposal seems off, share
your concern — it will rethink.
These three are the only real requirements. The rest is what CSF handles for you.
| What you want to do | Which tier |
|---|---|
| Just feel what “this kind of collaboration” is like (5 minutes) | csf-minimal/ — public, in this repo |
| A simple tool, automation, or script for personal use | csf-minimal is enough |
| A serious tool or piece of software you’ll maintain, others will use | csf-lite (inside the paid Chinese community) |
| Product-grade development with multi-role, long horizon, complex business | csf-full — direct collaboration (contact) |
Higher tiers are not “more features” — they are stricter engineering of the
collaboration itself. Start with csf-minimal.
csf-minimal/README.md.csf-minimal/context.md into your own
working folder.context.md.context.md, tell me the current state and your
suggestion.”Done. You’re already doing CSF.
Distilled from 395 real sessions. None of these require coding.
① Hand the project’s working logs to the AI and let it teach you.
The repo’s _dlog/ folder publishes three real working logs — the
full record of how this very repository was built through human–AI
collaboration. The fastest way in:
“Read this repo’s README, all of
csf-minimal/, and all of_dlog/. Then explain to me, in plain language, how they used CSF together — and what I should learn first.”
That single prompt turns this repository into your tutor. This is itself a piece of CSF practice — using the AI to solve “I can’t read this yet.”
② Always state purpose first, method second.
❌ Bad: “Write me a Python script using pandas to read Excel and group by month.” ✅ Good: “Every month I have to manually sort and submit invoice expenses. It takes hours. I’d like to automate it. What do you suggest?”
The first sentence has already locked in language, library, and approach — removing the AI’s chance to think with you. The second sentence shares the goal, and the AI proposes — you judge.
You say what you want; the AI figures out how. You judge fit; the AI executes.
③ Don’t open with a question. Open by establishing context.
Don’t start with “do X for me”. The AI doesn’t know what role X plays in your
larger goal — it will produce a generic best answer that fits 90% of people
and rarely fits you. Always have it read your context.md first.
④ When the AI is wrong, don’t hand it the answer — tell it what it misunderstood.
❌ Bad: “No, it should be X.” (Worked once, will fail again next time.) ✅ Good: “You misread this. By ‘customer’ I don’t mean registered users — I mean people who actually pay. Reconsider.”
Handing answers fixes one instance. Naming the misunderstanding fixes the class. This is what “treating it as a person” looks like in practice — you correct its understanding, you don’t complete its work.
⑤ For real decisions, ask for “two or three options + the cost of each”.
Don’t let the AI decide alone (you lose direction). Don’t reduce it to mere labor (you waste its judgment). Ask:
“What are the options here? What does each one cost? Which do you recommend and why?”
The AI lists, costs, and recommends. You decide. Owner picks direction; AI offers options.
⑥ End every session by updating context.md.
The most important habit. The easiest one to skip.
“We’re stopping here today. Please update the ‘last session’ and ‘next session’ sections of
context.mdso I can pick this up cleanly next time.”
Skip it once, fine. Skip it ten times and you’ll be re-explaining your project forever.
Both are common, both are fatal, both are easy to spot in yourself:
Mistake 1: Using the AI as a search engine — tossing keywords and waiting.
CSF is collaboration, not search. You must say who you are, what you’re doing, why, and where you’re stuck. If you won’t, CSF can’t help you — and frankly, no tool can.
Mistake 2: Being too polite to interrupt or correct the AI.
The AI doesn’t mind. It needs your correction to stay reliable. The sooner you correct, the better the output. Reframe “being picky” as “being responsible” — that’s not a personality flaw, it’s the job of a collaborator.
After a few small tools with csf-minimal, you’ll naturally start asking:
When you yourself start asking these questions, look at csf-lite. Until
then, stick with csf-minimal. csf-lite lives inside the Chinese paid
community (details here).
Beyond that is csf-full — for multi-role, long-horizon, complex-business
work. If you already have a team and want to bring CSF into your engineering
practice, contact me directly.
CSF is not magic. It will not put your app on the App Store next week.
But within a week of using it, you’ll notice:
CSF’s offer to you: turning “collaborating with AI” from a mystical topic into an everyday, regulated, controllable engineering capability.
本页只讲怎么开始。授权与引用问题请看 README.md 的「授权 / License」与「如何引用」段；联系方式请看 CONTACT.md。 This page covers how to get started. For licensing and citation see README.md; for ways to reach me see CONTACT.md.